Fusing appearance and distribution information of interest points for action recognition

Topic-Based Knowledge Transfer Algorithm for Cross-View Action Recognition

“…The action is represented by a complementary combination of local and global features, which are BoVW representation based on interest point descriptors [6] and shape-flow descriptors [14], respectively. Our low-level features are BoVW representation based on interest point descriptors [7]. [6] uses local information within a small region and tends to generate spurious detection in background areas, which affects the performance of the low-level action descriptors of [4], [12] and [13].…”

“…However, it uses local information within a small region and tends to generate spurious detection in background areas. In this paper, we adopt the algorithm proposed by [7], which overcomes the shortcomings of the Dollar detector.…”

Capturing relative motion and finding modes for action recognition in the wild

“…Accuracy Schuldt et al [9] 71.72% Dollar et al [11] 81.2% Reddy, Shah [22] 89.79% Fathi and Mori [37] 90.5% Klaser et al [18] 91.4% Marszalek et al [12] 91.8% egonzio et al [19] 94.33% Kovashka et al [27] 94.5% Gilbert et al [14] 94.5% RMD 91.2% RMD + Outlier Detection 94% RMD + Mode Finding 92.1% Table 2: Average accuracies on the KTH dataset using the Training/Validation/Test split defined in [9].…”

“…These include local jet descriptors [9], vector of concatenated pixel gradients [11], generalisation of the SIFT and SURF descriptors [17,18,16], and the HOG/HOF descriptors [12]. The approach by Bregonzio et al [19] differs significantly from the existing interest point based representation in that only the global distribution information of interest points is exploited. The detected interest points are typically then used in a discriminative [9,11] or generative [10] model.…”

3D R Transform on Spatio-temporal Interest Points for Action Recognition

“…All the feature vectors produced by different approaches are concatenated to form a larger feature vector. Liu et al [13], Ye et al [14] and Bregonzio et al [15] employ the kernel-level fusion approach and utilize a multi-kernel classifier for combining different features. The above fusion approaches rely only on the pairwise similarities of videos without considering the highorder correlations among videos.…”